AI-Proctored Written Exams Increase Dartmouth Final Grades by 0.71–1.30 SD

An experiment was run at Dartmouth College Spring 2026 on the Phosphor digital learning platform which integrates AI graded written quizzes into course readings. 90.2% of 151 students (3 sections of Introductory Statistics) voluntarily used the platform (instructors report readings generally only 10–15% of the time). Completing all 24 lessons and passing the three module reviews was associated with a gain on the final exam of 0.71 to 1.30 standard deviations relative to the control group.

Phosphor used Claude Sonnet 4.6 to automatically grade constructed-response (open-ended writing) questions using rubric scores defined by the instructor. The paper finds a dramatic effect of question type: the dosage–performance correlation vanished when the platform began using only multiple choice quizzes due to student feedback (R² = 0.001), but returned when written questions were added back into the third module. This suggests that the causal agent behind educational benefits is AI-graded writing assessments, rather than simply using AI tools. The results were published at the iTextbooks’26 workshop.